Reproduce result on Android Control datasset.

[SummerFall1819]

hello, thank you for your brilliant work! Recently we are conducting an experiment which requires to reproduce your results on Android Control. I write a script based on your inference.py file and the result is not very promising. I'd like to clarify the following issues.

1. which mode did you use during your inference? Because when I use default self plan mode, the model would consistently output mobile.open_app() throughout the trajectory despite the image.
2. How did you format your previous_actions? Your comment is slightly different from the Android Control Action Space and i'd like to know if you just use the original actions field from Android Control dataset.

Looking forward to hearing from you soon, thank you again for your consideration!